Independent component analysis in signals with multiplicative noise using fourth-order statistics

The existence of multiplicative noise greatly limits the applicability of independent component analysis (ICA), because it does not take into account the existence of the noise. This paper proposes a method to extend ICA to this kind of noisy environment, without any limitation in the nature of the sources or the noise. In order to do this, the statistical structure of a linear transformation of the noisy data is studied up to fourth order, and then this structure is used to find the inverse of the mixing matrix through the minimization of a cost function. The method designed is able to extract the mixing matrix and some statistical features of the noise and the sources, notably improving the performance of the standard ICA methods when the mixture is contaminated by multiplicative noise.